Barcelona, 06 May 2015 s2dverification Seasonal to decadal forecast verification in R Overview Nicolau Manubens

2 Outline – Introduction – Module diagram – Example of use – BigData issues – BigData approaches

3 Introduction – Forecast verification: evaluating model performance by comparing its output with observational data Origin (V. Guemas) On CRAN v2.3.2 – s2dverification (seasonal to decadal verification) is an R package that gathers various forecast verification tools coded by scientists, aiming to score models which run forecasts from a few seasons to few decades long.

4 Module diagram LOCAL STORAGE ESGF NODE or OPeNDAP SERVER s2dverification package BASIC STATISTICS SCORES Correlation, ACC, RMSSS, CRPS,... PLOTS PLOT DATA RETRIEVAL Fetches experimental and observational data sets stored in multiple conventions: NetCDF, NetCDF4 File per member, file per starting date,... Supports specific folder and file naming conventions Most time consuming: SCORES and DATA RETRIEVAL Anomaly Correlation Coefficient. 10M Wind Speed ECMWF S4 1 month lead with start dates once a year on first of November and Era-Interim in DJF from 1981 to Simple bias correction with cross-validation.

5 Example of use data <- Load('tas', exp = c('EnsEcmwfSeas', 'EnsMetfrSeas'), obs = c('ERAint'), nmember = 9, sdates = paste(1979:2005, '0501', sep = “”), leadtimemin = 2, leadtimemax = 4, latmin = 35, latmax = 75, lonmin = -25, lonmax = 70, storefreq = 'monthly', output = 'lonlat') Pick monthly 2-meter air temperature in J-J-A over Europe from ECMWF and Meteofrance ensemble experiments and from ERA-interim observation, from may 1st starting dates from 1979 to 2005.

6 Example of use meanModEnsSeas <- Mean1Dim(Mean1Dim(data$mod, 4), 2) meanModObsSeas <- Mean1Dim(Mean1Dim(data$obs, 4), 2) Average model and observational data along J-J-A (4th dimension) for all starting dates, and then average all models and observations across ensemble members. corr <- Corr(meanModEnsSeas, meanModObsSeas) Calculate the time correlation between each grid point in the model data and in the observational data. intervals <- seq(-1, 1, length.out = 21) PlotEquiMap(corr[1, 1,, ], data$lon, data$lat, brks = intervals) PlotEquiMap(corr[2, 1,, ], data$lon, data$lat, brks = intervals) Plot a map of correlations in between each experiment's and ERA-interim's data averaged in JJA and across ensemble members.

7 Example of use ECMWF ensemble experiment and ERA-interim's 2-m temperature averaged along JJA and across members correlation in Meteofrance ensemble experiment and ERA-interim's 2-m temperature averaged along JJA and across members correlation in

8 BigData issues – Computing time can raise to several hours in score computation or data retrieval. – Involved data occupies in some cases far more than the available main memory and hangs the machine. Example case: 2 x 9 x 27 x 3 x 17 x 39 x 8 bytes →7.7 Mbyte Usual case: 5 x 9 x 27 x 60 x 73 x 144 x 8 bytes → 6.1 Gbyte Big case: 1 x 50 x 36 x 120 x 144 x 288 x 8 bytes → 71.6 Gbyte ( n. of datasets x n. of members x starting dates x lead-times x latitudes x longitudes )

9 BigData approaches – In progress Avoid R memory duplications as much as possible. Wrappers of Fortran functions are always faster. Exploit multi-core. – Future work Store active data in disk instead of main memory. Parallelize verification tools to use on cluster. Reduce R's default representation precision.

For further information please contact Thank you!